The Toolkit Claude Code Kit Open Source · MIT Licensed 21 Commands · 14 Subagents · 7 Hooks github.com/stefanvekta/claude-code-kit Practice What We Preach The Toolkit Claude Code Kit Open Source · MIT Licensed 21 Commands · 14 Subagents · 7 Hooks github.com/stefanvekta/claude-code-kit Practice What We Preach
Reference Cheatsheet
The Toolkit·Open Source Release
MIT Licensed·Vekta·April 2026
§ The Toolkit

The kit we actually use.

A thoughtful CFO asks us — "Do you use this yourself, or are you just selling the idea?" Fair question. This is the answer. Our Claude Code setup. Genericized from a year of production work. Open-sourced because the patterns generalize.

Slash Commands
21 · Composable
Subagent Specialists
14 · Fresh Context
Deterministic Hooks
7 · 100% Enforced
License
MIT · No Wall
§ 01 · Why It Exists

Practice what we preach.

We've hired a lot of consultants. The good ones showed receipts. The bad ones were fluent in frameworks they didn't live by. So when we started telling clients to adopt AI seriously, we needed receipts of our own.

§The Receipts

This is what Stefan uses every day when he ships code at Vekta. A year of daily use on a production codebase before we genericized it.

MIT licensed. Clone it, install it, use it. No gate. No email wall. No "contact us for pricing." The code is the marketing.

The patterns generalize to any serious Claude Code setup. If it's useful to you, it's useful to you. If it sparks a conversation, better.

Origin Built on Yield, a production financial-modeling app. Genericized April 2026.

License MIT. Fork, modify, sell derivatives. No attribution required.

Maintenance Updated when we learn things. Roughly weekly.
§ 02 · The Opinion

One opinion. Three principles.

The difference between a senior engineer and a hobbyist isn't coding ability. It's judgment about blast radius. Seniors know what's reversible. They don't skip reviews. They don't push to main at 6 PM on a Friday. AI agents lack that judgment. The kit's job is to encode it mechanically — so the wrong thing is the hard thing to do.

01§ Principle
Writer / Reviewer separation

Split the doing from the grading.

No agent grades its own output. Writer and reviewer, always. The author subagent writes. A separate reviewer subagent in a fresh context grades. Sycophancy is structural. We solve it with structure.

Implementationcode-reviewer, security-reviewer, verification-auditor. Fresh context. None see the author's reasoning.
02§ Principle
Spec-driven delivery

Specs are contracts.

Every plan step ties to a verifiable criterion. Nothing is "done" until an independent auditor confirms it — with file:line citations. If an item can't be cited, it isn't complete.

Gate/verify-plan blocks the PR if any plan item lacks file:line evidence. No exceptions.
03§ Principle
Session health discipline

Session health is non-negotiable.

Context above 70% degrades. Three wrong answers in a row is a smell. Two hours elapsed is time to checkpoint. The best session ends before it degrades.

SignalStatus line shows context%, model, branch, cost. Red past 70% — that's the signal to /checkpoint.
§ The Thesis
The difference between senior and hobbyist is judgment about blast radius. The kit makes that judgment structural — so the wrong thing is the hard thing to do.
House position · the whole reason this kit exists
§ 03 · The Surfaces

Four configuration surfaces.

Claude Code exposes four places where behavior can be shaped: memory, hooks, commands, subagents. Each does a different job. Picking the right one is the difference between rules that fire and rules that feel good in a document.

a.§ Surface
Advisory · ~80% adherence

Memory. CLAUDE.md + rule imports.

What Claude should know on session start. Project shape, conventions, pitfalls. Advisory — about eighty percent adherence. Good for taste and workflow rules. Not the right place for safety rules.

FilesCLAUDE.md, .claude/rules/stack.md, .claude/rules/kit-discipline.md.
b.§ Surface
Deterministic · 100% enforced

Hooks. Seven shell scripts.

Shell commands firing on every lifecycle event. Exit code 2 blocks the tool call and shows the error back to Claude. 100% deterministic. Blocks secrets in diffs, force-pushes, destructive rm, wrong package-manager, push-to-main, broken JSONB joins.

Filesblock-secrets · protect-bash · forbid-dangerous-patterns · format-on-save · session-start-banner · status-line · context-warn.
c.§ Surface
Composable · User-invoked

Commands. Twenty-one slash commands.

Short prompts registered as /name. /plan, /sharpen, /review, /security-review, /ship-it, /verify-plan, /checkpoint. Each does one thing. Together they form the gate stack.

Used daily/sharpen, /plan, /review, /ship-it, /pr.

Used weekly/deps-update, /eval-review, /a11y-audit.
d.§ Surface
Specialists · Fresh context

Subagents. Fourteen specialists.

Named sub-Claudes with narrow jobs — security-reviewer, verification-auditor, supervisor, visual-validator, prompt-auditor, db-migration-expert. Fresh context. The thing that makes writer/reviewer separation structural.

Opus-tiersecurity-reviewer, db-migration-expert, verification-auditor, supervisor, visual-validator, prompt-auditor.

Inheritthe rest.
§ 04 · The Gate Stack

The gate stack.

Every real change runs through this sequence. Seven gates. Each catches a different failure mode. Skip them and the kit is decoration. Run them and you get professional-engineer discipline without the professional engineer.

01 · Specify
/sharpen
Scores a rough ask 0-10. Unlocks planning only when specific enough.
02 · Plan
/plan
Enters plan mode. Produces plan with risks and rollback. Waits for approval.
03 · Build
supervisor
Orchestrates specialists in parallel. Synthesizes into one go/no-go decision.
04 · Verify
/verify-plan
Fresh-context auditor cites file:line evidence. Blocks if incomplete.
05 · Review
/review
Principal-engineer read of the diff. Security, data-loss, N+1, accessibility.
06 · Pre-Ship
/ship-it
Pre-deploy GO/NO-GO: tests, secrets, rollback path, feature-flag defaults.
07 · Open PR
/pr
Opens a draft PR only if pr-describer and verification-auditor both approve.
§ 05 · Install

Three commands.

One-time per machine. Per-project thereafter. No frameworks. No build step. No account. Idempotent — existing files get backed up with timestamps before the kit overwrites.

# 1. Clone the kit
git clone https://github.com/stefanvekta/claude-code-kit \
  ~/claude-code-kit
# 2. One-time, per machine
~/claude-code-kit/setup.sh --user-level
# 3. Per project
cd ~/code/your-project
~/claude-code-kit/setup.sh
§ The Reference

Pop it out. Dock it next to your desk.

The cheatsheet is 540 pixels wide on purpose. Mono font. One page. Every command, every subagent, every hook — in the order you'll reach for them. Pop it out of the browser, resize skinny, park it beside Claude Desktop. Leave it there.

  1. 01Click the button. A new tab opens.
  2. 02Drag that tab out to its own window.
  3. 03Resize it narrow — around 540px wide.
  4. 04Dock it beside Claude Desktop. Forget about it.
Open the Cheatsheet
§ Why We Open-Sourced It
We didn't invent writer/reviewer separation. We just wired it up so you can't skip it.
Vekta · open-sourced April 2026 · github.com/stefanvekta/claude-code-kit
§ 06 · Limitations

What it deliberately doesn't do.

The most important section. If you're evaluating whether to install this, read this carefully.

01§ Limitation

Not a beginner tutorial.

Assumes you know JSONB, can read a Bash script, have shipped production code. If you're new to Claude Code, start with the official docs. A scaffold, not an onboarding.

Start insteaddocs.claude.com/claude-code.
02§ Limitation

Not a productivity hack.

It adds gates. The first time /sharpen scores your ask 3/10, it'll feel slower. It is slower — by design. The trade is fewer catastrophic failures.

Rough honestPer-task time: +20%.
Blast-radius failures: -90%.
03§ Limitation

Opinionated. May disagree with yours.

We hard-block pnpm, push to main, certain JSONB joins, writes to env files. Clone and rewrite the pieces you disagree with — that's what MIT is for.

Opinions baked innpm over pnpm · no push to main · writer/reviewer · fresh-context verify.
04§ Limitation

Not magic.

The kit enforces discipline. It catches what a tired human misses at midnight. That's the whole value. If you already have staff-engineer habits, the marginal gain is small.

Marginal valueHighest for senior operators shipping solo.
05§ Limitation

Not a replacement for judgment.

Claude is still wrong sometimes. The gates catch a lot; they don't catch everything. Read the diffs. The kit helps — it doesn't substitute. Anyone saying otherwise is selling.

The contractIf you can't explain every line, don't ship it.
§ 07 · Fit

Who this is for.

A good fit.
Say Yes If This Is You

Senior engineers shipping production code with Claude. Non-engineers fluent enough to read code and run real apps. Small teams that want the discipline without writing the config from scratch.

Probably not.
Say No If This Is You

First-time Claude Code users. Teams that want "AI magic" without changing their workflow. Anyone who would override the hooks the first time one blocked them.

§ Next Step

Clone it. Or talk to us.

Going to run it? Clone the GitHub repo. Running AI adoption at a professional-services firm and any of this resonates? Let's talk. The engagement we do most of is the AI Fluency rollout →.

First conversation is free. Ninety minutes. You walk away with a clearer read on your organization either way.

Phone
+46 76 118 18 15
Location
Stockholm, Sweden